Time series data, like your emotional state or the relentless noise of the internet, is often messy. It's full of random fluctuations, sudden spikes, and dips that obscure the underlying patterns. To see the "big picture" – the general trend or the smoother cycles – we can use techniques to filter out this noise. One of the simplest and most common methods is the moving average.
A moving average smooths out the data by taking the average of a fixed number of data points at a time, and then "moving" that window of points forward. Instead of looking at a single day's screen time, you look at the average over the past 3 days, or 5 days, or whatever window size you choose. This process dampens the effect of extreme values and short-term variations, revealing the smoother, underlying movement.
Algorithms use smoothing techniques like this to make your behavior appear more predictable. By focusing on your average engagement over a period, they can better estimate what you're likely to do next, ignoring the messy, unpredictable details of your individual moments of rebellion (or despair).
Let's apply a moving average to our hypothetical happiness data and see how it smooths out the dramatic ups and downs.
The first step is to choose a window size ($w$). This is the number of consecutive data points you will average at a time. A larger window size results in a smoother line but can lag behind recent changes.
In our visualization, you can select a window size using the dropdown menu.
For each data point $y_t$ in the time series (starting from $t=w$), the moving average $MA_t$ is calculated as the average of the current data point and the previous $w-1$ data points.
$MA_t = \frac{y_{t-w+1} + y_{t-w+2} + \dots + y_t}{w} = \frac{1}{w} \sum_{i=t-w+1}^{t} y_i$
For example, with a window size of $w=3$ and data $Y = \{8, 7, 6, 7, 5, 4, 6, 5, 4, 3\}$:
$MA_3 = \frac{y_1 + y_2 + y_3}{3} = \frac{8 + 7 + 6}{3} = \frac{21}{3} = 7$
$MA_4 = \frac{y_2 + y_3 + y_4}{3} = \frac{7 + 6 + 7}{3} = \frac{20}{3} \approx 6.67$
And so on, for each subsequent data point. The moving average series will have $w-1$ fewer points than the original series.
The red line in the visualization shows the calculated moving average for the selected window size, smoothing out the blue line of the original data.
The resulting moving average series provides a smoothed version of the original data, making it easier to identify the underlying trend and longer-term cycles by reducing the impact of short-term noise.
While it sacrifices some detail (those dramatic spikes and drops are averaged away), the moving average gives algorithms a clearer signal of your general behavior, making their predictions and content recommendations more "accurate" within the confines of your predictable, smoothed-out existence.